SpaceX vs. Blue Origin: The Fight Over Data Centers in Space
SpaceX vs. Blue Origin: The Fight Over Data Centers in Space
In early March 2026, Amazon asked federal regulators to reject SpaceX's plan to fill low Earth orbit with a million computers. Thirteen days later, Blue Origin — the rocket company owned by Amazon's founder, Jeff Bezos — filed its own application to launch 51,600 of them. The chairman of the FCC, in between, publicly criticized Amazon over its own satellite deployment record rather than answering the petition procedurally.
That sequence is the most revealing thing that has happened in this industry all year, and almost nobody wrote it down in one place. It tells you what these companies actually believe, as opposed to what they say in press releases.
The short version:
- Two applications, wildly different scales. SpaceX asked the FCC for up to 1,000,000 orbital data center satellites. Blue Origin asked for 51,600. Both are chasing the same market: AI compute demand that Earth's power grid is struggling to supply.
- Amazon tried to kill the SpaceX filing, calling it incomplete and unrealistic — then Bezos's other space company filed a proposal resting on the same unproven premise thirteen days later.
- Neither has been approved. Both asked to be excused from the FCC's standard deployment deadlines.
- Then its rocket exploded. A New Glenn blew up on the stand during a May 28 ground test, taking out the one pad it can launch from. It has not flown since.
- The hard constraints apply to both equally. Launch cost and heat rejection constrain the two companies identically, and neither has published numbers showing they close.
A few months ago, I got into an argument with a friend about Elon Musk's plan to put data centers in orbit. My position was straightforward: in a vacuum, thermal radiation is the only way to shed heat, which seemed brutally inefficient at data-center scale. My friend's reply was immediate: "Do you really think he'd announce this without knowing that?"
We were both arguing about physics. Reading what has happened since, I think we picked the wrong thing to argue about. The physics is hard but tractable. What will actually decide this is a launch price, a regulator, and whether either company can keep its hardware working long enough to matter.
Two Filings, Seven Weeks Apart
SpaceX moved first. On January 30, 2026, it applied for authority to launch and operate the Orbital Data Center System, docket SAT-LOA-20260108-00016, which the FCC Space Bureau accepted for filing on February 4. The application covers up to one million satellites between 500 and 2,000 kilometers, in 30-degree and sun-synchronous inclinations, inside orbital shells up to 50 kilometers thick, linked primarily by optical inter-satellite lasers. SpaceX described the system in the filing as a first step toward a civilization able to harness the full output of its star.
Blue Origin filed on March 19 for Project Sunrise: up to 51,600 satellites in sun-synchronous orbits between 500 and 1,800 kilometers, in planes of 300 to 1,000 spacecraft, relaying data through its planned TeraWave broadband constellation. Same thesis, same orbital band, about one-twentieth the scale.
Both companies asked the FCC to waive its standard milestone rules, which normally require half of an authorized constellation deployed within six years and the remainder within nine. SpaceX's own waiver argument describes what those deadlines are for: preventing operators from warehousing spectrum they cannot actually use. My read is that when both applicants for a new class of infrastructure ask to be released from the anti-warehousing test, that tells you something — though a waiver request on a first-of-its-kind system is not by itself evidence of bad faith.
The one-million figure also needs a caveat. It is a ceiling in an authorization request, not a construction commitment, and constellation filings routinely ask for more headroom than the operator intends to use.
The Fight: Amazon, the FCC Chairman, and Thirteen Awkward Days
On March 6, Amazon's low Earth orbit division formally petitioned the Commission to deny the SpaceX application, describing it as incomplete, speculative, and unrealistic. Its substantive argument was not unreasonable: SpaceX had offered little detail on how a million spacecraft would maneuver, avoid each other, and be disposed of, and approving that at once would hand one operator enormous influence over the orbital shells everyone else has to fly through.
Amazon summarized the filing as "a lofty ambition rather than a real plan".
Five days later, FCC Chairman Brendan Carr answered publicly rather than procedurally, posting that Amazon might better spend its energy on the thousand-odd satellites it was running short of its own deployment milestone. Whatever you make of a regulator arguing with a petitioner on social media while their application is pending, it told the industry something about the reception the SpaceX filing was getting.
Then the argument got awkward. Blue Origin's Project Sunrise application landed on March 19, from the rocket company Amazon's founder owns. SpaceX wrote to the FCC the next day to make the obvious point: whatever standard the Commission applies to a speculative orbital compute constellation ought to apply to Blue Origin's as well.
What that sequence suggests: Amazon's stated objection was that nobody can responsibly build orbital data centers at this scale yet. Thirteen days later, Bezos's rocket company filed to build them. Both positions can be sincere — 51,600 is genuinely more defensible than a million, and Amazon's orbital-safety concerns are real ones. But the timing means the petition can also be read as a fight over who gets there first, not only over whether anyone should.
The docket drew close to 1,500 comments, according to the American Astronomical Society, which filed its own concerns about night-sky brightness and radio interference. As of early September 2026, no public decision on either application has been reported.
Then Blue Origin Lost Its Launch Pad
Just over two months after filing for 51,600 satellites, Blue Origin's ability to launch anything at all took a serious hit.
New Glenn's third flight, on April 19, 2026, went two ways at once. The first stage came home and landed on the recovery ship, a real milestone given that the booster was on its second flight. The upper stage did not cooperate. One of its two engines came up short on the second burn, and the AST SpaceMobile spacecraft it carried ended up in an orbit it had no way to climb out of. The customer wrote the satellite off. The FAA opened a mishap investigation and did not clear the rocket to fly again until May 22.
Six days after that clearance, on the evening of May 28, engineers lit the first stage's seven BE-4 engines for a static-fire ahead of the next flight. The rocket blew up on the stand. The fireball took the whole vehicle, booster and fueled upper stage together, and tore through Launch Complex 36 — the only pad New Glenn can fly from. One of the pad's lightning towers came down. Nobody was hurt, and the 49 Amazon Leo satellites the flight was meant to carry were still on the ground.
The company traced the cause to one of those BE-4 engines, and Limp said in early August that the trouble started at the engine's main oxygen valve, a finding confirmed by recovered hardware. Dave Limp had reported in early June that the propellant tanks and water tower came through better than feared, sparing Blue Origin the replacements with the longest manufacturing lead times, and that the company was aiming to fly again before the end of the year.
An earlier version of this article said SpaceX was the only serious entrant at constellation scale. That was true when it was written and stopped being true in March, and by June the balance had moved again. Any comparison of these two companies written before the summer needs updating, including this one.
| Company | What it filed for | Strongest position | Weakest link |
|---|---|---|---|
| SpaceX Orbital Data Center System / Starmind |
Up to 1,000,000 satellites, 500–2,000 km; NVIDIA Vera Rubin hardware; first launch targeted Q4 2027 | Direct control of the launch vehicle, satellite manufacturing, the optical network, and an integrated AI workload through xAI | Thermal and radiation engineering at a scale never demonstrated; its own IPO prospectus says 2028 at the earliest |
| Blue Origin Project Sunrise |
Up to 51,600 satellites, 500–1,800 km sun-synchronous; filed March 19, 2026; relays via TeraWave | A far smaller ask than SpaceX's, backed by an owner who can fund it without public markets | Grounded since the May 28 pad explosion; NG-3's upper stage underperformed in April; the TeraWave relay layer it depends on does not exist yet |
| Amazon Leo (formerly Project Kuiper) |
3,236-satellite broadband constellation; not an orbital compute system | Deep AWS integration, committed capital, and satellites actually reaching orbit | No onboard compute architecture at all, running behind its own FCC deployment milestone, and its New Glenn launches are now delayed too |
| D-Orbit / Loft Orbital | Existing players in small-scale on-orbit edge computing and hosted payloads | Operating real compute hardware in orbit today | Built for modest onboard processing rather than the continuous high-power workloads frontier AI inference requires |
The competitive picture cuts both ways. That two of the best-capitalized space companies on Earth independently filed for orbital compute constellations within seven weeks suggests the thesis is not fringe. That one of them spent the summer rebuilding a launch pad instead of launching is a reminder of how much sits between a filing and a working constellation.
The Number Both Companies Need and Neither Has
Strip away the corporate theater and both applications rest on the same arithmetic. Lifting hardware to orbit has to get cheap enough that a satellite full of GPUs competes with a building full of them.
Google's Project Suncatcher research used roughly $200 per kilogram to low Earth orbit as the point where the comparison starts to become interesting, and other orbital-compute studies have adopted similar figures. It is worth being clear about what that number is: a modeling assumption used to test whether the concept survives contact with arithmetic, not a market price and not an established break-even. Real viability also depends on payload density, spacecraft lifetime, utilization, radiator mass, and replacement cadence.
Against that, there is exactly one Starship price that exists as a signed commercial contract rather than a projection. Voyager Technologies disclosed in its annual filing that it will pay $90 million to launch the Starlab station on Starship. Starlab is a roughly 65-tonne spacecraft. Divide the one by the other and Voyager is paying on the order of $1,400 per kilogram for what it is actually putting in orbit — about seven times the modeling threshold.
You can be more generous to Starship than that. Spread the same $90 million across the 100 to 150 tonnes the vehicle is credited with lifting, as though the station filled the bay, and the figure drops to somewhere between $600 and $900 per kilogram. That is the best case the contract supports, and it is still three to four times the number the business case wants.
Why that gap matters: it is the difference between a spreadsheet that closes and one that does not, and every projection about this industry lives inside it. Be careful what the number is, though. It is not a Starship price list. It is one disclosed contract for one unusual payload, divided two different ways — which is precisely why it is worth showing both.
To be fair to the optimistic case, that contract prices a dedicated launch of an awkward, low-density station in 2029, before Starship reaches routine cadence. Satellites shipped by the thousand in an optimized dispenser are a different economic animal. The honest version is not that $200 is impossible; it is that the only contracted number anyone can point to today sits several times above it, however generously you divide it.
The same arithmetic is harsher for Blue Origin. It has no vehicle in Starship's class flying, and after May its cadence is zero. The charge Amazon leveled at SpaceX, that the ambition had run ahead of the plan, is hard to square with his other company's own position.
Heat: The Constraint That Ignores Both of Them
Every watt a chip consumes becomes a watt of waste heat that has to leave. On Earth that happens through fans, chilled water, or heat exchangers. In vacuum none of those exist. There is no air, no convection, no surrounding mass to absorb the load. Heat still moves inside the spacecraft by conduction, but the only way off the vehicle is thermal radiation.
A commonly cited first-order estimate puts one megawatt of waste heat at around 1,200 square meters of radiator area, or about 1.2 square meters per kilowatt. Apply that to the announced hardware. Reported power figures for SpaceX's first satellite have ranged around 120 to 150 kilowatts, which lands the radiator somewhere between 145 and 180 square meters. That is the floor area of a modest house, unfolded beside a single satellite, doing nothing but throwing heat away.
I want to be careful about what that calculation is. It uses a generic figure, not either company's design, and the real number swings hard with radiator operating temperature, coating emissivity, view factor, and how much sunlight and Earth infrared the panel has to fight. A well-optimized design could come in meaningfully under it.
Neither SpaceX nor Blue Origin has published radiator assumptions, which is exactly why an outside estimate is useful: it tells you the order of magnitude the design has to beat.
The full thermal analysis: a companion piece works through the radiator physics in detail, including why published estimates for the same one-megawatt load range from under 1,000 square meters to over 3,000.
Read: Why AI Data Centers in Space Will Fail Without Giant Radiators →
A spacecraft thermal radiator panel. Radiators get far less attention than solar arrays, but they are one of the hard limits on sustained compute in orbit, scaling directly with the heat load — for either company.
Where SpaceX Actually Is Now
On August 24, 2026, Musk posted that SpaceX and NVIDIA had designed a space-optimized version of the Vera Rubin NVL72 system for launch in the fourth quarter of 2027, with meaningful scale in 2028. The first satellite is called Starmind AI1. He described the design, in remarks reported by Bloomberg, as "significantly simpler, lower cost, denser and lighter than a traditional rack."
Set that against what the same company told the SEC. In the prospectus that took it public, SpaceX described orbital compute deployment beginning as early as 2028 and warned, in risk language quoted by Reuters, that the initiative may not achieve commercial viability. I have not read the prospectus itself; that wording comes secondhand from the reporting. Both statements can be defensible — one is a legally cautious floor, the other an engineering target — but the S-1 language is the one written under liability, and it is the more conservative of the two.
The money arrived earlier in the year. Reuters and CNBC reported on February 2 that SpaceX had acquired xAI in an all-stock deal valuing the combined company near $1.25 trillion. On June 12 SpaceX listed on the Nasdaq as SPCX at $135 a share, raising $75 billion; three days later its underwriters exercised the overallotment option in full, bringing total proceeds to about $85.7 billion — the largest IPO ever completed.
The merger supplied the workload; the offering supplied the capital. Blue Origin has no comparable public-market financing; its funding runs through Bezos.
What Neither Filing Fully Answers
Radiation is the first gap. Cosmic radiation causes bit flips, memory faults, and processor instability at rates that would be intolerable in a terrestrial facility. The traditional answer is radiation-hardened silicon, but hardened parts run generations behind the commercial accelerators that make frontier inference economically viable. A space-optimized Vera Rubin variant is SpaceX's stated answer; what that optimization costs in performance has not been made public. (The platform is named for the astronomer whose observations established the case for dark matter — her story is worth reading.)
Collision risk at these densities
SpaceX had launched close to 13,000 Starlink satellites by early September 2026, with roughly 9,800 of them operational. The Secure World Foundation, in its own filing on Blue Origin's application, put the total active low Earth orbit population at around 15,000. A million-satellite authorization is about seventy times that number. Astronomers reviewing the filings have put the increase closer to a hundredfold, using their own baseline. Either figure describes a change of kind rather than degree.
Higher density raises both the routine burden of conjunction management and the risk of a debris-generating collision, which is the mechanism behind Kessler Syndrome — the cascade in which each impact makes the next more likely. How bad that gets depends on orbital altitude, debris lifetime, relative velocities, and how reliably a million spacecraft can maneuver.
The worrying case is not the gradual one: a sufficiently severe cascade could make particular orbital regimes unsafe to use for a very long time, and that is not something a later decision can undo.
What comes back down
When aluminum-bodied satellites burn up on reentry, they produce aluminum oxide, which collects in the stratosphere and mesosphere rather than falling out. A 2025 CIRES and NOAA modeling study in the Journal of Geophysical Research: Atmospheres found that reentry volumes consistent with expected constellation growth by 2040 could slow Southern Hemisphere polar vortex wind speeds by about 10 percent and produce temperature anomalies up to 1.5 degrees Celsius at high latitudes. The authors modeled the particles as chemically inert, so downstream ozone effects remain poorly characterized. Neither application resolves this, and no regulator currently requires an applicant to.
Verdict: Who Wins, and What "Winning" Would Require
SpaceX has the strongest vertically integrated position, and the gap is substantial. It controls the rocket, the satellite line, the laser network, and, since the xAI deal, the workload itself. Blue Origin has the rocket and the satellite line; the relay network does not exist yet, and there is no in-house AI workload to feed. Its launch vehicle has also been grounded since May.
Blue Origin still has the more approvable proposal. Fifty-one thousand satellites would still be an extraordinary authorization, far larger than anything the FCC has approved to date. But a million is the number that forces the Commission to decide what low Earth orbit is for.
Neither wins on physics or price. Both need launch costs well below the only contracted figure available, radiator structures that dwarf the compute modules they cool — on the estimates above, a house's floor area for a single satellite — and radiation-tolerant silicon that does not yet exist as a product line.
And the fight itself is a tell. When a competitor argues the technology is not ready and then files to build it thirteen days later, at least part of the disagreement is about market position rather than feasibility.
So who won the argument with my friend? Neither of us, I think. We were debating whether the heat could be managed, when the decisive variables turned out to be a price per kilogram, a regulator's patience, and whether a rocket survives a ground test. The heat has to go somewhere, and that is real, but it is an engineering problem, and engineering problems yield to money and time. The rest is schedule risk, and 2026 has been a good year for reminders about schedule risk.
If SpaceX's stated timeline holds, Starmind AI1 flies in late 2027 — one satellite carrying a rack's worth of compute, which is a demonstration, and SpaceX is very good at demonstrations. Whether the millionth one ever flies is a different question, and the honest answer is that it depends on a spreadsheet nobody outside these two companies has seen.
Frequently Asked Questions
Q. Why did Amazon try to block SpaceX's satellite data centers?
Amazon's satellite division asked the FCC in March 2026 to throw the application out. Its stated reason was that a constellation that size raised collision-avoidance and disposal questions SpaceX had barely addressed, and that granting it would hand one operator too much say over altitudes everyone shares. What complicated the argument was the calendar: Bezos's rocket company put in a competing orbital compute application of its own less than two weeks later. The Commission has not ruled on either.
Q. What is Blue Origin's Project Sunrise?
It is Bezos's answer to the SpaceX filing. Submitted to the FCC in March 2026, it seeks authority for as many as 51,600 compute satellites in sun-synchronous orbit, with the company's planned TeraWave network carrying their traffic. At roughly a twentieth of what SpaceX asked for, it is the smaller and more conventional of the two proposals — though it would still be the largest constellation any regulator has ever cleared.
Q. Is Blue Origin's New Glenn rocket still flying?
No. A static-fire test on May 28, 2026 ended with the rocket blowing up on the pad at Launch Complex 36, the only place New Glenn can launch from. Blue Origin later pinned the failure on the main oxygen valve of a BE-4 engine, and says it hopes to launch again within the year. The flight before that one, in April, recovered its booster but stranded a customer satellite in a useless orbit.
Q. When will the first data center satellites actually launch?
SpaceX is aiming for late 2027 with a single satellite, Starmind AI1. Its own IPO prospectus gave a later and softer answer — deployment starting no sooner than 2028, with a warning that the venture might never pay for itself. Blue Origin has named no date at all, and neither company holds an approved licence yet.
Sources & References
- FCC Space Bureau Public Notice DA 26-113 (February 4, 2026), accepting SpaceX's Orbital Data Center System application, ICFS File No. SAT-LOA-20260108-00016, filed January 30, 2026: up to one million satellites, 500–2,000 km, 30-degree and sun-synchronous inclinations, orbital shells up to 50 km, optical inter-satellite links.
- Secure World Foundation's filing on Blue Origin's orbital data center application for the current LEO population figure (around 15,000 active satellites) and for its argument that these applications represent a qualitative rather than incremental shift in constellation scale.
- SpaceNews on the January 30 filing, including SpaceX's request to waive the FCC's six-year and nine-year deployment milestones and its argument that those rules exist to prevent spectrum warehousing.
- Broadband Breakfast on Blue Origin's Project Sunrise application, filed March 19, 2026: up to 51,600 satellites, 500–1,800 km, sun-synchronous, relayed via TeraWave.
- The Register (March 9, 2026) on Amazon's petition to deny, and The Register (March 23, 2026) on SpaceX's response letter citing Blue Origin's filing.
- CNBC (March 11, 2026) on FCC Chairman Brendan Carr's public response to Amazon's opposition and Amazon's shortfall against its own deployment milestone.
- SpaceNews (May 28, 2026) and NASASpaceflight (May 2026) on the New Glenn static-fire explosion at Launch Complex 36, the loss of both stages, and the toppling of one launch tower; SpaceNews (June 2026) on Dave Limp's damage assessment and the end-of-year return-to-flight target; and Space.com on the 49 Amazon Leo satellites manifested for the cancelled June 4 flight. Root cause was announced in early August 2026 and is covered by Spaceflight Now (August 6, 2026), which reports Limp attributing the anomaly to the main oxygen valve on one of the BE-4 engines. FAA clearance to resume flights was granted May 22, following the April 19 NG-3 mission in which an upper-stage engine underperformed on its second burn.
- American Astronomical Society action alert, noting nearly 1,500 comments filed on the SpaceX docket and characterizing the constellation as roughly a hundredfold increase over the current LEO population.
- Light Reading and Bloomberg (August 24–25, 2026) on the Starmind timeline and the space-optimized NVIDIA Vera Rubin design. Reported power figures for the first satellite have varied across outlets and should be treated as provisional.
- Reuters analysis (April 2026) on orbital data center skepticism, the Microsoft undersea data center precedent, and SpaceX's S-1 risk language on commercial viability.
- Voyager Technologies Form 10-K (March 2026) disclosing a $90 million Starship launch for the Starlab station. Starlab's mass is given as roughly 65 tonnes dry by NextSpaceflight and by SpaceQ, which implies about $1,400/kg for the actual payload; dividing instead by Starship's quoted 100–150 t lift capacity gives $600–$900/kg. Analyst projections for fully reusable Starship at scale run lower, in the $100–200/kg range. All pricing figures are commercial estimates and may not reflect undisclosed contract terms. Google's Project Suncatcher research is the most commonly cited source for the ~$200/kg modeling threshold.
- EE Times on orbital heat rejection: approximately 1,200 m² of radiator area per megawatt at practical operating temperature. A first-order estimate governed by the Stefan–Boltzmann law, not a design specification; published estimates for the same load range from under 1,000 to over 3,000 m².
- CNBC (June 11, 2026) on the $135 IPO price and $75 billion raise, and CNBC (June 15, 2026) on the full greenshoe exercise lifting total proceeds to $85.7 billion; Reuters and CNBC (February 2–3, 2026) on the SpaceX–xAI merger.
- CIRES and NOAA modeling study (Maloney et al.), Journal of Geophysical Research: Atmospheres, 2025: alumina accumulation from satellite reentry, roughly 10 percent reduction in Southern Hemisphere polar vortex wind speed and up to 1.5 K temperature anomalies at high latitudes under 2040 constellation scenarios, with particles modeled as chemically inert.
How to read the numbers in this article
Three tiers. The two FCC applications, Amazon's petition, the chairman's response, the New Glenn explosion, the merger, and the IPO are confirmed events anchored to docketed filings and wire reporting. The radiator area and the per-kilogram launch figures are estimates: one a physics rule of thumb sensitive to temperature and emissivity, the other a single disclosed contract divided by a payload mass range. Launch dates, power budgets, and deployment scale are projections that have already moved and will move again. Verify against primary filings before drawing conclusions.
Comments
Post a Comment